Multiple imputation when records used for imputation are not used or disseminated for analysis

Abstract

When some of the records used to estimate the imputation models in multiple imputation are not used or available for analysis, the usual multiple imputation variance estimator has positive bias. We present an alternative approach that enables unbiased estimation of variances and, hence, calibrated inferences in such contexts. First, using all records, the imputer samples m values of the parameters of the imputation model. Second, for each parameter draw, the imputer simulates the missing values for all records n times. From these mn completed datasets, the imputer can analyse or disseminate the appropriate subset of records. We develop methods for interval estimation and significance testing for this approach. Methods are presented in the context of multiple imputation for measurement error.